Multi-Objective Optimization Using

Volume 2 No.
2 ISSN 2079-8407
Journal of Emerging Trends in Computing and Information Sciences
©2010-11 CIS Journal. All rights reserved.

http://www.cisjournal.org
Multi-Objective Optimization Using

Multi Parent Crossover Operators
Rahila Patel1, M.M.Raghuwanshi2
RCERT, Chandrapur, M.S., India
NYSSCoER, Nagpur,M.S., India
rahila.patel@gmail.com, m_raghuwanshi@rediffmail.com
ABSTRACT
The crossover operator has always been regarded as the primary search operator in genetic algorithm (GA) because it
exploits the available information from the population about the search space. Moreover, it is one of the components to
consider for improving the behavior of the GA. To improve performance of GA multi parent crossover operators have been
used. Multi parent crossover operators involve sampling of features of more than two parent solution into the offspring that
accelerated speed of convergence to global optima. These operators are based on some probability distribution and are
gene-level parent centric crossover operators. In this work, we have used MPX (multi-parent crossover with polynomial
distribution) and MLX (multi-parent cross-over with lognormal distribution) operators for multi-objective optimization.
The performance of these operators is investigated on commonly used multi-objective functions. GA used for
experimentation is Non-dominated Sort Genetic Algorithm-II (NSGA-II). It is observed that these operators work well
with NSGA-II and have given encouraging results.
Keywords—Multi-objective optimization, Non-dominated Sort Genetic Algorithm-II (NSGA-II), Crossover operator, MPX (multi-
parent polynomial distribution crossover), MLX (multi-parent lognormal distribution crossover).
I. INTRODUCTION elitist solutions can not participate in the genetic action

fully, although the elitist strategy had improved the
Real- world optimization problems generally efficiency greatly. To protect the diversity among
involve multiple objectives. Many researchers have solutions, multi parent re-combination operators have been
proved that Genetic Algorithm which is a randomized used in this work.
search algorithm inspired by principles of natural genetics In this work, we have used MPX and MLX
can solve real-world multi- objective optimization operators for multi-objective optimization. The
problems. Real-coded Genetic Algorithm (RCGA) uses performances of these operators are investigated on
real number vector as chromosome that makes it close to commonly used multi-objective optimization functions. It
the natural formulation of many problems. RCGA possess is observed that these operators work well with NSGA-II
several characteristics that are desirable to solve and have given encouraging results.
optimization problems In RCGA the recombination This paper is organized as follows. In Section 2,
operation is a method of sharing information among we discuss multi parent crossover operators. In Section 3,
chromosomes. The detailed study of many recombination we explain the NSGA-II algorithm. Experimentation in
operators can be found elsewhere [1][2]. section 4 and discussion on results have been reported in
Multi-parent recombination operator MPX and Section 5. Finally we conclude this paper in Section 6.
MLX are proposed by Raghuwanshi et al [10]. These
neighborhood-based, gene-level crossover operators are II. MULTI PARENT CROSSOVER
multi-parent extension of simulated binary crossover OPERATOR FOR REAL CODED
operator (SBX)[11] and SBX with lognormal distribution
(SBX-l)[8] respectively. These operators define a
GA
probability distribution of offspring solutions based on
some measure of distance among the parent solutions. Usually, the crossover operator is applied to pairs
Depending upon the current level of diversity in the of chromosomes, generating two offspring for each one of
population they may favor the production of additional them, which are introduced in the population. However,
diversity (divergence) or the refinement of the solutions multiparent crossover operators have been proposed,
(convergence). which combine the features of more than two parents for
The NSGA-II presented by Deb et.al. [6] has generating the offspring. In general; sampling of more in
been widely accepted by many researchers around the formation from a population helps evolution process to
world for its better diversity and faster convergence in bring better changes in the next generation. Beyer and Deb
solutions. However, the NSGA-II also has its shortage, [5] and Kita et al. [4] have given guidelines to design
that is, with the decrease of diversity among all solutions, recombination operator. Recombination operators used in
search rates of overall solution lowers and local Pareto RCGA are based on uniform, polynomial, triangular or
solutions is arrived prematurely because most of non- lognormal probability distribution. Eiben has done very
99
Volume 2 No. 2 ISSN 2079-8407

good study on multi-parent operators used in evolutionary a. From population select best parent and pick other
computing [12]. The exploitative and explorative behavior (µ-1) solutions randomly.
of op-erators due to use of probability distribution and b. For each gene (i=1,n) in real-parameter
more sampling of information due to use of more than two chromosome execute following steps :
parents are the strength of multi-parent recombination
operators.[7]
The parent-centric recombination operator i) Choose ui randomly from the interval [0,
preserve the population means by producing off-spring 1].
near the parent and assigns each parent an equal ii) Compute βi using (4).
probability of creating offspring in its neighborhood. The
−zη
PCX [15], SBX [11] and SBX-l [8] are parent-centric e if u
i ≤ 0 .5
β  (4)
recombination operators. [7] i
zη
II.I Multi-Parent Polynomial e Otherwise
Distribution Crossover (MPX) Operator Where z ~ N (0, 1) is standard normal variable.

iii) Calculate D using (2)
The MPX operator is a multi-parent extension of iv) Generate two genes around gene of best
the SBX operator.The prototype algorithm for the MPX parent (say x1) using 3
operator is as follows:
a. From population select best parent and pick other Operators with lognormal distribution are more
(µ-1) solutions randomly. explorative i.e. it is capable to generate genes away from
b. For each gene (i=1,n) in real-parameter chro- the parent gene. Their exploration range increases with
mosome execute following steps increase in η. Also their probability of creating genes near
the parent gene is almost zero.[10]
1 /(η  1)
(2u i ) if u i ≤ 0.5
βi  1 /(η  1) (1) III. NON-DOMINATED SORTING
Otherwise GENETIC ALGORITHM (NSGA-II)
1 /( 2(1 − u i )) 
The Fast Elitist Non-dominated Sorting Genetic
Algorithm for Multi-Objective Optimization, NSGA-II
1 successfully combines the following key elements:
yi x i 1 (βi * D ) (3)
1. A fast-non-dominated sorting approach.
2. A density estimator.
  k 3. A crowded comparison operator.
D  ( ∑ ( ∑ | x j − x |)/)/
(2)
k1 j1 i i In the algorithm’s main loop, a random parent
population Po is initially created. The population is sorted
i) Choose ui randomly from the interval [0, 1]. based on non-domination. Each solution is assigned a
ii) Compute βi using (1). fitness equal to its non-domination level (1 is the best
iii) Calculate level). Thus, minimization of fitness is assumed. Binary
iv) Generate two genes around gene of best parent tournament selection, recombination, and mutation
(say x1) using (3) operators are used to create a child population Qo of size
N [6]. Next repeat the following steps.
The operators based on polynomial distribution
are more exploitative and exploitation range decreases Step1: Combine parent and offspring population and
with increase in distribution index of probability create Rt= Pt Performs a non-dominated sorting to Rt and
distribution (η).[10] identify different fronts Fi ,I=1,2,…, etc
Step 2: Set new population Pt 1 φ Set a counter I=1,
II.II Multi-parent Lognormal Until | Pt 1 |+ Fi < N , perform Pt 1= Pt 1 Fi and I= I+1.
distribution Crossover (MLX) operator Step 3: Perform the crowding sort Fi , c procedure and
include the most widely spread N − Pt 1  solution by
The MLX operator is a multi-parent extension of using the crowding distance values in the sorted Fi to Pt 1
SBX-l operator. Step 4: Create offspring population Qt 1 from Pt 1 by
using the crowded tournament selection, crossover and
A prototype algorithm for the MLX operator is as follows: mutation operators. NSGA-II has made use of SBX
100

crossover operator for recombination. In this work we TABLE I PARAMETER SETTING USED
have used MPX and MLX crossover operator for re- IN THE STUDY
combination.
Parameter Values
IV. EXPERIMENTATION
GA type NSGA-II
Table I shows parameter setting used for Population size (N) 100
experimentation. Test problems used in the study are Number of generations 250
given in the Appendix. These test problems are chosen
Crossover probability pa-
from a number of significant past studies as mentioned in 0.9
rameter (pc)
[3, 13, 14]. All problems have two objective functions and Probability Distribution in-
all objective functions are to be minimized. 1 - 10 in step of 1
None of these problems have any constraint. In dices for crossoverηc
this work the performance indicator used to quantify the Probability Distribution in-
1-20 in step 0f 1
quality of the obtained results is the GD (Generational dices for mutationηm
Distance) and IGD (Inverted Generational Distance) Probability of variable 0.5
metric given in the appendix.
change (pvc)
Disruptiveness of chromosome depends upon the
number of genes changed due to recombination operation. Number of Parents 5
This disruptiveness on the one hand leads to more diverse Number of children (λ) 2
exploration that can prevent premature convergence, but
on the other hand, it slows convergence speed at the same V. DISCUSSION ON RESULTS
time. For both the operators the value of cross-over
probability is moderated i.e. pvc=0.5.We use the 30 random simulations are performed for each problem.
polynomial mutation with mutation probability of pm  All the results are taken after 250 generations. Table II
0.1. shows the mean and variance of the Generational Distance
(GD) obtained using SBX, SBX-l, MPX and MLX
operators with real coded NSGA-II for multi-objective
functions KUR, ZDT1, ZDT2, ZDT3, ZDT4 and ZDT6.
Multi parent operators MPX and MLX have converged
better in all problems, as com-pared to SBX and SBX-l.
MPX and MLX have shown convergence to Pareto
optimal front in all problems except ZDT4.
Table II Mean GD and Variance GD of Different Multi-objective functions with SBX, SBX-L,
MPX and MLX operators.
NSGA-II-SBX NSGA-II-SBX-l NSGA-II-MPX NSGA-II-MLX

Function
Mean Var.GD Mean Var.GD Mean Var.GD Mean Var.GD
ZDT1 0.008345 0.001448 0.009028 0.001473 0 0 0 0
ZDT2 0.012074 0.002965 0.01322 0.00307 0 0 0 0
ZDT3 0.008209 0.001400 0.00805 0.00133 0 0 0 0
ZDT4 0.90470 6.33890 0.7380 5.433 0.00057 0.000001 0.00043 0.000001
ZDT6 0.056955 0.018227 0.05608 0.01839 0 0 0 0
KUR 0.02896 0.000018 ------- ------- 0 0 0 0
The minimum (Min), maximum (Max), mean, and V.I Simulation results with MPX Operator
standard deviation (S.D.IGD) of the IGD metric are
reported in Table III and IV. The IGD metric measures
We have graphically shown the results obtained by
both the convergence and the spread of the obtained
solutions. Smaller the value of the IGD metric, better is NSGA-II-MPX with parameters pc =0.9, pvc=0.5, ηc =1
the obtained solution. All the IGD values are close to zero. and ηm =20 for problems ZDT2 and ZDT3. These figures
demonstrate the abilities of multi parent crossover
operator to help in converging to the true front and in
finding diverse solutions in the front.
101

TABLE III IGD

METRIC FOR MPX
Func-
MinIGD MaxIGD MeanIGD S.D.IGD
tions
ZDT 0.2789168 0.02461812
1 0.062225397 0.28284271 6 0
ZDT 0.2739953 0.04589630
Fig 2 Non-dominated Solutions with NSGA-II-MPX on
2 0. 0 0.28284271 9 5
ZDT3
ZDT 0.2745950 0.03776283
3 0.014142136 0.28284271 2 5
Fig 2 shows non-dominated solutions for ZDT3
ZDT 0.08708727 0.01733937
function.
4 0.050911688 0.13293607 1 0
The Pareto-optimal front is discontinuous and
ZDT 0.2669017 0.05989974
6 0. 0 0.28284271 0 9 NSGA-II-MPX has got success in finding all the
K U 0.00138592 0.01142876 discontinuous regions with uniform spread of non-
R 0. 0 0.050911688 9 0 dominated solutions.
The exploitative nature of MPX operator has
produced a search-bias that has helped the algorithm to
TABLE IV IGD METRIC converge better in all problems. MPX operator has helped
FOR MLX to find a better.
Fun- MinIGD MaxIGD MeanIGD S.D.IGD

V.II Simulation results with MLX Operator
tions
ZDT1 0.05939697 0.2728427 0.2758282 0.03017187 Simulation results of NSGA-II-MLX for test
ZDT2 0.0 0.2758427 0.2643561 0.06733055 problems ZDT4 and ZDT6 is shown in fig 3 and 4.
ZDT3 0.01131370 0.2018427 0.2651933 0.05394973 Function ZDT4 requires more generations to converge to
ZDT4 0.01697056 0.01838477 0.05509776 0.04678049 Pareto-optimal front. Around 400 generations are required
ZDT6 0.0 0.2828427 0.2355174 0.09693670 for convergence near to
K 0.0 0.0113 0.0011 0.009
UR 1370 3137 751943
The non-dominated solutions on the problem ZDT2 is

shown in Fig 1.
Fig 3 Non-dominated Solutions with NSGA-II-MLX on

ZDT4
Fig 1 Non-dominated Solutions with NSGA-II-MPX on

ZDT2
This problem has a non-convex Pareto- optimal Fig4. Non-dominated Solutions with NSGA-II-MLX on
front. NSGA-II-MPX has found a better spread and more
solutions in the entire Pareto-optimal region. ZDT6
102

Pareto-optimal front. MLX operator has given Algorithms behavior is shown graphically in fig (e)-(h).
excellent performance for ZDT6 problem. Fig 4 shows ZDT1 is a simple problem and the algorithm convergences
Pareto-optimal front, which is non-uniform very nicely with MLX operator. Since MLX is explorative
operator and it exploration (capability to produce genes
away from parent genes) in-creases with increase in
probability distribution index ηc . Higher explorative
power has produces gene very much away from the parent
genes and hence there are less number of non-dominated
solution in the Pareto front. Poor performance of NSGA-
II-MLX with higher values of ηc is seen in fig (e)-(h).
Therefore we suggest keeping value of ηc 1 or 2.
CONCLUSIONS
In this work we have tested multi-parent parent-centric

crossover operators, MPX and MLX for multi-objective
functions. NSGA-II-MPX and NSGA -II -MLX are able
to maintain a better spread of solutions in the obtained
non-dominated fronts and shown good converge for multi-

Fig (a)-(d) shows effect of different values of distribution
index for MPX crossover on Pareto-optimal front and non-
dominated solutions. Values of ηc are a) 2, b) 4, c) 6, d) 8
with very good density of non-dominated solutions on the

Pareto optimal front. Lognormal distribution is explorative
in nature hence MLX operator produces offspring away
fro m parent. Also it samples more than two parent, more
diverse solutions it produces which helps in a better search
of decision variable space. NSGA-II-MLX has given good
performance with all the six problems.
V.III Effect of Probability Distribution

Index for crossover ηc
We have tested MPX and MLX operators for

parameter probability of distribution index ηc . For MPX
operator we have selected parameter pc =0.9, pvc=0.5, ηm
=20 and ηc = 2 to 8in step of 2. Fig a)-d) Shows simulation
results for MPX operator for ZDT1 problem. It is observed
that with increase in the value of probability distribu-tion
index for MPX crossover η c , the Pareto-optimal front
deteriorates i.e. discontinues with less number of non-
dominated solutions in Pare-to front and convergence is
also weak. The rea-son is the exploitation capability of
MPX that decreases with the increase η c . Hence the study
Fig (e)-(h) show effect of different values of distribution
suggests keeping value of ηc in the range of 1-4. in-dex for MLX operator. Values of ηc are e) 2, f) 3, g)4,
MLX operator with parameter pc =0.9, pvc=0.5, ηm =1 and h) 5
objective test problems. MPX and MLX opera-tors have
ηc = 2 to 5 in step of 1 is tested on ZDT1 function. improved performance of the algorithm as is shown in
103

table 3 and 4. [6] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A

The empirical result shows that, the use of multiple parent fast and elitist multi-objective genetic algorithm:
recombination operators has pro-duced good results. We NSGAII,” IEEE Trans. on Evolutionary Computation
conclude with: 6 (2), pp. 182-197, April 2002.
 As mentioned earlier the polynomial distribution [7] M.M. Raghuwanshi and O.G. Kakde, “Probability
is more exploitative and lognormal distribution is distribution based recombination operator to solve
more explorative in nature. The exploitative unimodal and multi-modal problems”, International
nature produces search-bias suitable to solve Journal of Knowl-edge-based and Intelligent
unimodal function and explorative nature Engineering Systems 9, 1–9. 2006
produces search-bias that help operators to
overcome number of local minima in multi- [8] M.M. Raghuwanshi, O.G. Kakde, P.M. Singru, U.
modal functions.  Kale, “Simulated Binary Crossover with Lognormal
 Divergence and convergence of solution is Distribution”, In Proceedings of the 7th Asia-Pacific
Conference on Complex Systems (Complex 2004) 6-
controlled by the calculation of D (average of 10 Dec. 2004.
average distance between the parent solutions)
that marked MPX and MLX operators self- [9] M.M. Raghuwanshi and O.G. Kakde, “Multi-parent
adaptive. Recombination Operators with Multiple Probability
 Use of more than two parent solutions helps Distributions for Real Coded Genetic Algorithm”,
operator to sample more features into off-spring 10th on-line world conference on soft computing in
that accelerated speed of convergence to global industrial application (WSC10) 19th sep-8th oct,
Pareto optimal front.  2005
Future work will be on multi parent operator with mixed [10] M. M. Raghuwanshi and O. G. Kakde, “Multi-parent
probabilities and self-adaption of multi parent operator for Re-combination operator with Polynomial or
various multi-objective optimization problems. Lognormal Distribution for Real Coded Genetic
Algorithm” 2nd Indian International Conference on
Artificial Intelligence (IICAI), pp. 3274-3290, 2005.
REFERENCES
[1] Deb, K., Multi-Objective Optimization using [11] Deb K. & Agrawal R.B., Simulated binary crossover
Evolutionary Algorithms, John Wiley & Sons, New for continuous search space, Complex System 9,
York. 2001 115-148 ,1995
[12] Eiben, A.E., “Multi -parent recombination”, In T.

[2] Herrera, F., Lozano, M., Sánchez, A.M. A taxonomy
Bäck, D.B. Fogel, and Z. Michalewicz, editors,
for the crossover operator for real-coded genetic
Evolutionary Computation 1: Basic Algorithms and
algorithms. An experimental study. International
Operators, pages 289-307, Institute of Physics
Journal of Intelligent Systems 18(3): 309-338.2003
Publishing, 2000.
[3] Deb K., “Multi-objective genetic algorithms: Problem
difficulties and construction of test functions,” in [13] E. Zitzler, K. Deb, and L. Thiele, “Comparison of
Evol. Com-put., vol. 7, pp. 205–230. 1999 multi-objective evolutionary algorithms: Empirical
results,” Evol. Comput., vol. 8, no. 2, pp. 173–195,
[4] Kita, H., Ono I. and Kobayashi, S., “Multi-parent Summer 2000.
extension of the Unimodal Normal Distribution
Crossover for real-coded genetic algorithm”, In proc. [14] F. Kursawe, “A variant of evolution strategies for
of the International conference on evolutionary vector optimization, “in Parallel Problem Solving
computation’99 (IEEE press, Pis-cataway, New from Nature, H.-P. Schwefel and R. Männer, Eds.
Jersey, pp. 646-651, 1999. Berlin, Germany: Springer-Verlag, pp. 193–197.
1990.
[5] Beyer H.G. and Deb K., “Self-Adaptive Genetic
Algorithms with Simulated Binary [15] K.Deb, A. Anand and D. Joshi, “A computationally
Crossover”,Technical Report No. CI-61/99, effi-cient evolutionary algorithm for real parameter
Department of Computer Science/XI, University of optimization”, Evolutionary Computation Journal
Dortmund, Germany, 2001. 10(4): 371-395, 2002.
104

APPENDIX:
1) Performance measures: 2) Test Problems
Generational Distance (GD) is given by: Problem Objective functions

KUR n −1
x
f1 ( x)  ∑i 1 (−10 exp( −0.2
2 2
i  xi 1 ))
n 0 .8
f (x)  (x
∑
3
2 i 1 i  5 sin x i )
The parameter di is the Euclidian distance ZDT1 f1(x)  x1
between the solution Qi and the nearest member of P*. f ( x )  g ( x ) 1 −  
2 x1 / g ( x )
Where P* is Pareto-optimal set and Q is non-dominated n
solution set. g ( x )  1  9 (∑ i  2 x i ) /( n − 1 )
Inverted Generational Distance(IGD): The IGD
metric meas-ures “how well is the Pareto-optimal front
ZDT2 f1(x)  x1
represented by the obtained solution set”. To quantify this f 2 ( x )  g ( x ) 1 − ( x 1 / g ( x ))  2

information, a large set of evenly spaced points on the n
g ( x )  1  9 ( ∑ i  2 x i ) /( n − 1 )
Pareto-optimal front is generated. Let the size of this set
be H. The mini-mum Euclidean distance of each point in ZDT3
f1(x)  x1
this set from the obtained solution set is computed. Let
x1
this distance be li for the ith element of the Pareto-optimal f 2 ( x )  g ( x ) 1 − x1 / g ( x ) − sin( 10 πx1 )
set. Then the IGD meric is given by g(x)

n
g ( x )  1  9 ( ∑ i  2 x i ) /( n − 1)
ZDT4 f1(x)  x1
f2 (x)  g(x) 1 − x1 / g(x)  
n 2
g ( x )  1  10 ( n − 1 )  ∑ i  2 x i − 10 cos( 4 π x i )
ZDT6 f 1 ( x )  1 − exp( − 4 x 1 ) sin 6 (6πx1 )

f 2 ( x )  g ( x ) 1 − ( f 1 ( x ) / g ( x ))
2

n 0 . 25
g (x) 1 9 ∑ i2 xi /( n − 1 ) 
105

Multi-Objective Optimization Using

Transféré par

Informations du document

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

Multi-Objective Optimization Using

Transféré par

Droits d'auteur :

Formats disponibles

Volume 2 No.

©2010-11 CIS Journal. All rights reserved.